Algebraic Adversarial Attacks on Integrated Gradients

TL;DR

This paper introduces algebraic adversarial examples to generate adversarial attacks on integrated gradients, providing a mathematically tractable method to understand vulnerabilities in explainability models for neural networks.

Contribution

The paper proposes a novel algebraic approach to create adversarial examples for integrated gradients, highlighting new vulnerabilities in explainability methods.

Findings

Algebraic adversarial examples can be systematically generated for integrated gradients.

The approach reveals specific conditions under which explanations can be manipulated.

Provides a new mathematical framework for studying adversarial attacks on explainability methods.

Abstract

Adversarial attacks on explainability models have drastic consequences when explanations are used to understand the reasoning of neural networks in safety critical systems. Path methods are one such class of attribution methods susceptible to adversarial attacks. Adversarial learning is typically phrased as a constrained optimisation problem. In this work, we propose algebraic adversarial examples and study the conditions under which one can generate adversarial examples for integrated gradients. Algebraic adversarial examples provide a mathematically tractable approach to adversarial examples.